In-memory compute (IMC), also called processing-in-memory, performs some computation inside the memory itself, reducing the amount of data shuttled between separate memory and processing units.
This matters because moving data typically costs more energy and time than the computation it feeds. AI workloads make the problem acute: neural networks involve huge arrays of weights and repeated matrix operations, so cutting data movement directly improves power efficiency and latency.
ReRAM (RRAM) is a candidate for IMC because it stores data as resistance states. In some architectures those resistance states can represent neural-network weights, so computation happens within the memory array itself. This is the basis of analog compute-in-memory and neuromorphic computing research.
The nearer-term architecture is near-memory compute (NMC). Embedding fast, low-power non-volatile memory on the same chip as microcontrollers or AI accelerators keeps code, weights and data close to the processor, reducing reliance on external memory.
Weebit ReRAM is relevant in both timeframes. Today, embedded ReRAM supports near-memory compute in edge AI and embedded systems. It offers a foundation for architectures where the memory array participates directly in computation.